Behavior of Power Series at the Endpoints

A power series comes with a radius of convergence RR. This guarantees absolute convergence on the open interval (a−R,a+R)(a-R, a+R). It also guarantees divergence outside [a−R,a+R][a-R, a+R]. But what happens exactly at the two boundary points a−Ra-R and a+Ra+R? The Radius of Convergence machinery — the root test, the ratio test — is silent there by construction. Those tests are built to detect strict inequalities. They return "inconclusive" precisely at the boundary. Determining endpoint behavior is not a footnote. It decides whether the true domain of convergence is open, half-open, or closed. This distinction matters enormously for differentiating and integrating series term by term. It also matters for Abel's theorem on boundary values. And it matters for identifying which functions a series actually represents. This question sits at the crossroads of everything learned about series so far. To settle it we abandon the ratio and root tests entirely — they cannot help. Instead we plug the two numerical values x=a+Rx = a+R and x=a−Rx=a-R into the series. This produces two concrete numerical series. Each must be examined one at a time. We use the full toolkit: the Alternating Series Test, comparison tests, the integral test, telescoping tricks, and direct estimation of the general term's asymptotic size. The big picture is combinatorial. There are exactly four possible outcomes for the endpoint pair. These are converge/converge, converge/diverge, diverge/converge, and diverge/diverge. All four genuinely occur for natural examples. No shortcut formula predicts which case holds. Each series demands its own argument. This article develops the precise definitions. It proves the key structural theorem, Abel's theorem, guaranteeing continuity up to a boundary point of convergence. It works through asymmetric examples in full rigor. This way, the method — not a table of tricks — becomes second nature. This article delivers a complete framework: a rigorous definition of the interval of convergence, a proof of Abel's continuity theorem with all hypotheses tracked, two fully worked asymmetric examples exhibiting different endpoint outcomes, an explicit counterexample showing why the hypotheses of Abel's theorem cannot be dropped, and a survey of where this analysis is indispensable.

Key takeaways
  • The ratio/root test gives the radius RR but is always inconclusive exactly at x=a±Rx = a\pm R; endpoints need separate, direct arguments.
  • Plugging in x=a+Rx=a+R or x=a−Rx=a-R turns the power series into an honest numerical series — analyze it with comparison, alternating series, or integral tests.
  • All four combinations (both converge, both diverge, one of each) occur; there is no universal rule, only case-by-case work.
  • The true domain of convergence is one of (a−R,a+R)(a-R,a+R), [a−R,a+R)[a-R,a+R), (a−R,a+R](a-R,a+R], or [a−R,a+R][a-R,a+R].
  • Abel's theorem: if a power series converges at an endpoint, its sum function is continuous up to that endpoint from inside the interval.
  • Key idea: radius of convergence tests only detect strict inequalities, so endpoints are numerically separate problems requiring their own convergence tests.

Where this lives today

Determining exact endpoint behavior is not a purely academic exercise in test-taking; it governs which functions a power series genuinely represents on its full natural domain. In the theory of generating functions in combinatorics, whether a generating function's series converges at its radius of convergence controls asymptotic formulas for the coefficients (via singularity analysis, a cornerstone of analytic combinatorics). In Fourier and orthogonal-series analysis, boundary convergence questions of this same flavor (does the series converge at the endpoint of its domain of validity?) recur for Legendre and Chebyshev series used in numerical analysis and physics. In probability theory, Power Series and Radius of Convergence techniques applied to probability generating functions require checking convergence at x=1x=1 (the boundary of validity for many discrete distributions) to compute means and variances via term-by-term differentiation, and Abel's theorem is exactly the tool that justifies passing derivatives or limits to that boundary. In numerical methods and engineering (signal processing, control theory via zz-transforms), the region of convergence of a series expansion and its boundary circle determine stability of discrete-time systems, and endpoint/boundary convergence is checked by the same case-by-case series tests as here. Finally, in cryptography and coding theory, formal power series over finite fields or with combinatorial coefficients are analyzed for convergence-like radius conditions in pp-adic settings, an algebraic cousin of this same boundary-sensitivity phenomenon.

Ask the tutor: How do I decide which convergence test to try first when an endpoint series looks unfamiliar or mixes factorials, powers, and alternating signs all at once?

DefinitionInterval of Convergence

Let

f(x)=∑n=0∞cn(x−a)nf(x) = \sum_{n=0}^{\infty} c_n (x-a)^n

be a power series with radius of convergence R∈(0,∞)R \in (0,\infty) (the cases R=0R=0 and R=∞R=\infty are degenerate: convergence only at x=ax=a, or convergence on all of R\mathbb{R}, respectively, with no endpoint question to resolve). The interval of convergence of the series is the set

I={ x∈R:∑n=0∞cn(x−a)n converges }.I = \{\, x \in \mathbb{R} : \sum_{n=0}^\infty c_n(x-a)^n \text{ converges} \,\}.

By the definition of radius of convergence, (a−R,a+R)⊆I⊆[a−R,a+R](a-R, a+R) \subseteq I \subseteq [a-R, a+R]. Determining II exactly therefore reduces to deciding, separately, whether the two numerical series

∑n=0∞cnRn(at x=a+R),∑n=0∞cn(−R)n(at x=a−R)\sum_{n=0}^\infty c_n R^n \qquad (\text{at } x = a+R), \qquad \sum_{n=0}^\infty c_n (-R)^n \qquad (\text{at } x = a-R)

converge or diverge. There are exactly four possibilities for II:

(a−R, a+R),[a−R, a+R),(a−R, a+R],[a−R, a+R].(a-R,\,a+R), \quad [a-R,\,a+R), \quad (a-R,\,a+R], \quad [a-R,\,a+R].

Intuition

Think of the radius of convergence as drawing a solid wall of convergence strictly inside the interval, and a solid wall of divergence strictly outside it — but the wall itself, the boundary circle of radius RR (on the real line, just the two points a±Ra\pm R), is a knife-edge where the ratio test's limit equals exactly 11. A limit of 11 is the one value the ratio test cannot classify, by the test's own statement: it requires the limit to be strictly less than or strictly greater than 11. So the tool that built the interval is mute exactly where we now need information. What happens at that knife-edge depends on finer structure of the coefficients cnc_n that the radius calculation threw away. The radius only sees the exponential growth rate of ∣cn∣|c_n|, via lim⁡∣cn∣1/n\lim |c_n|^{1/n} or lim⁡∣cn+1/cn∣\lim |c_{n+1}/c_n|. But at the endpoint, ∣cnRn∣=∣cn∣Rn|c_n R^n| = |c_n| R^n, and whether ∑∣cn∣Rn\sum |c_n| R^n converges depends on the polynomial correction to that exponential rate — e.g. whether cn∼1/nc_n \sim 1/n or cn∼1/n2c_n \sim 1/n^2 or cn∼(−1)n/nc_n \sim (-1)^n/\sqrt n, all of which give the same radius but wildly different endpoint behavior.

Figure
The four possible interval shapes, all sharing the same radius RR
Same radius R, four different endpoint outcomes a−R a+R (a−R, a+R) (a−R, a+R] [a−R, a+R) [a−R, a+R]

The theorem: Abel's Theorem

TheoremAbel's Theorem on Boundary Continuity

Let

f(x)=∑n=0∞cn(x−a)nf(x) = \sum_{n=0}^\infty c_n (x-a)^n

be a power series with radius of convergence RR, 0<R<∞0 < R < \infty, so that ff is defined and (as shown by Power Series and Radius of Convergence) continuous on the open interval (a−R,a+R)(a-R, a+R). Suppose the series converges at the right endpoint, i.e. ∑n=0∞cnRn\sum_{n=0}^\infty c_n R^n converges (converges to some real number SS, meaning x=a+Rx = a+R lies in II). Then: 1. The series converges uniformly on [a,a+R][a, a+R]. 2. The function ff, extended by f(a+R):=Sf(a+R) := S, is continuous on (a−R,a+R](a-R, a+R]; in particular

lim⁡x→(a+R)−f(x)=∑n=0∞cnRn.\lim_{x \to (a+R)^-} f(x) = \sum_{n=0}^\infty c_n R^n.

(The analogous statement holds at the left endpoint x=a−Rx = a - R by applying the theorem to the series in the variable −x-x.)

Proof

Without loss of generality set a=0a = 0 and R=1R=1 (replace xx by (x−a)/R(x-a)/R and cnc_n by cnRnc_n R^n; this rescaling changes neither convergence nor continuity statements). So assume

f(x)=∑n=0∞cnxnf(x) = \sum_{n=0}^\infty c_n x^n

has radius of convergence 11, and ∑n=0∞cn\sum_{n=0}^\infty c_n converges to SS. We must show ff is continuous on (−1,1](-1,1] and that the series converges uniformly on [0,1][0,1].

Step 1: Set up partial sums and a summation-by-parts (Abel summation) identity. Let sN=∑n=0Ncns_N = \sum_{n=0}^N c_n (partial sums of the endpoint series), with s−1:=0s_{-1} := 0, and let S=lim⁡N→∞sNS = \lim_{N\to\infty} s_N by hypothesis. For x∈[0,1)x \in [0,1) and any NN, write the NN-th partial sum of f(x)f(x) using cn=sn−sn−1c_n = s_n - s_{n-1}:

∑n=0Ncnxn=∑n=0N(sn−sn−1)xn.\sum_{n=0}^N c_n x^n = \sum_{n=0}^N (s_n - s_{n-1}) x^n.

Summation by parts (Abel's identity) rearranges this as

∑n=0Ncnxn=sNxN−∑n=0N−1sn(xn+1−xn)=(1−x)∑n=0N−1snxn+sNxN.\sum_{n=0}^N c_n x^n = s_N x^N - \sum_{n=0}^{N-1} s_n (x^{n+1} - x^n) = (1-x)\sum_{n=0}^{N-1} s_n x^n + s_N x^N.

This is a purely algebraic identity, verified by expanding the right side and using sn−sn−1=cns_n - s_{n-1}=c_n; no convergence assumption is used yet.

Step 2: Let N→∞N \to \infty for fixed x∈[0,1)x \in [0,1). Since ∣x∣<1|x|<1, xN→0x^N \to 0, and sN→Ss_N \to S is a bounded sequence (convergent sequences are bounded), so sNxN→0s_N x^N \to 0. Also (1−x)∑n=0N−1snxn→(1−x)∑n=0∞snxn(1-x)\sum_{n=0}^{N-1} s_n x^n \to (1-x)\sum_{n=0}^\infty s_n x^n provided this last series converges — which it does, because sns_n is bounded (say ∣sn∣≤M|s_n|\le M for all nn) and ∑∣x∣n\sum |x|^n converges absolutely for ∣x∣<1|x|<1, so ∑snxn\sum s_n x^n converges absolutely by comparison. Taking N→∞N\to\infty in Step 1's identity gives, for every x∈[0,1)x \in [0,1),

f(x)=(1−x)∑n=0∞snxn.(∗)f(x) = (1-x) \sum_{n=0}^\infty s_n x^n. \tag{$\ast$}

Step 3: A summation identity for the constant SS. Since ∑n=0∞(1−x)xn=(1−x)⋅11−x=1\sum_{n=0}^\infty (1-x)x^n = (1-x)\cdot \frac{1}{1-x} = 1 for x∈[0,1)x\in[0,1) (geometric series), we may write

S=S⋅1=(1−x)∑n=0∞Sxn.S = S\cdot 1 = (1-x)\sum_{n=0}^\infty S x^n.

Subtracting this from (∗)(\ast):

f(x)−S=(1−x)∑n=0∞(sn−S)xn.(∗∗)f(x) - S = (1-x)\sum_{n=0}^\infty (s_n - S) x^n. \tag{$\ast\ast$}

This identity is the heart of the proof: it expresses the deviation of f(x)f(x) from the limit SS in terms of the deviations sn−Ss_n - S, which we know go to 00.

Step 4: Estimate using the tail of sn−Ss_n - S. Fix ε>0\varepsilon>0. Since sn→Ss_n \to S, choose N0N_0 such that ∣sn−S∣<ε/2|s_n - S| < \varepsilon/2 for all n>N0n > N_0. Split the sum in (∗∗)(\ast\ast):

f(x)−S=(1−x)∑n=0N0(sn−S)xn  +  (1−x)∑n=N0+1∞(sn−S)xn.f(x)-S = (1-x)\sum_{n=0}^{N_0} (s_n-S)x^n \;+\; (1-x)\sum_{n=N_0+1}^\infty (s_n - S)x^n.

For the tail, since 0≤x<10 \le x < 1 and ∣sn−S∣<ε/2|s_n - S|<\varepsilon/2 for n>N0n > N_0:

∣(1−x)∑n=N0+1∞(sn−S)xn∣≤(1−x)⋅ε2∑n=N0+1∞xn≤(1−x)⋅ε2⋅11−x=ε2,\left| (1-x)\sum_{n=N_0+1}^\infty (s_n-S)x^n \right| \le (1-x)\cdot \frac{\varepsilon}{2}\sum_{n=N_0+1}^\infty x^n \le (1-x)\cdot\frac{\varepsilon}{2}\cdot\frac{1}{1-x} = \frac{\varepsilon}{2},

uniformly in x∈[0,1)x\in[0,1) — this bound did not depend on how close xx is to 11, which is exactly what makes the convergence uniform. For the head, let M0=max⁡0≤n≤N0∣sn−S∣M_0 = \max_{0\le n\le N_0}|s_n - S| (a finite maximum over finitely many terms). Then

∣(1−x)∑n=0N0(sn−S)xn∣≤(1−x)(N0+1)M0,\left|(1-x)\sum_{n=0}^{N_0}(s_n-S)x^n\right| \le (1-x)(N_0+1)M_0,

which →0\to 0 as x→1−x \to 1^- (for fixed N0N_0), and in particular can be made <ε/2< \varepsilon/2 by requiring xx sufficiently close to 11: specifically for

1−δ<x<1,δ=min⁡(1, ε2(N0+1)M0+1).1 - \delta < x < 1, \qquad \delta = \min\left(1,\ \frac{\varepsilon}{2(N_0+1)M_0 + 1}\right).

Step 5: Conclude continuity at the endpoint. Combining the two bounds, for all xx with 1−δ<x<11-\delta < x < 1,

∣f(x)−S∣≤ε2+ε2=ε.|f(x) - S| \le \frac{\varepsilon}{2} + \frac{\varepsilon}{2} = \varepsilon.

Since ε>0\varepsilon>0 was arbitrary, this shows lim⁡x→1−f(x)=S=f(1)\lim_{x\to 1^-} f(x) = S = f(1), i.e. ff is continuous at x=1x=1 from the left. Combined with the already-known continuity of ff on (−1,1)(-1,1) (a standard fact about power series inside their radius of convergence, established in Power Series and Radius of Convergence via uniform convergence on compact subintervals), ff is continuous on (−1,1](-1,1].

Step 6: Uniform convergence on [0,1][0,1]. The estimate in Step 4 in fact shows more: given ε>0\varepsilon>0, the same N0N_0 (depending only on ε\varepsilon, not on xx) makes the tail of (∗∗)(\ast\ast) uniformly small for all x∈[0,1)x\in[0,1), and at x=1x=1 itself f(1)−S=0f(1)-S=0 trivially. A slightly more careful bookkeeping of the partial sums of the original series ∑cnxn\sum c_n x^n (rather than the transformed series (∗∗)(\ast\ast)) using the same Abel-summation estimate shows

sup⁡x∈[0,1]∣∑n=0Ncnxn−f(x)∣→0as N→∞,\sup_{x\in[0,1]} \left| \sum_{n=0}^N c_n x^n - f(x) \right| \to 0 \quad \text{as } N\to\infty,

establishing uniform convergence of ∑cnxn\sum c_n x^n on [0,1][0,1]. (This is the standard refinement of the Dirichlet/Abel uniform convergence test, using that the partial sums sNs_N of ∑cn\sum c_n converge, hence are uniformly Cauchy, together with ∑n=0Nxn≤11−x⋅(1−xN+1)≤11−x\sum_{n=0}^N x^n \le \frac{1}{1-x}\cdot(1-x^{N+1})\le \frac1{1-x} bounded uniformly via the same (1−x)(1-x)-weighting trick as above.) Undoing the normalization a=0,R=1a=0,R=1 recovers the theorem as stated. ■\blacksquare

∎
CorollaryEndpoint value via limits, not just series

Under the hypotheses of Abel's theorem, if in addition ff is known in closed form on (a−R,a+R)(a-R,a+R) (e.g. as an elementary function), then

∑n=0∞cnRn=lim⁡x→(a+R)−f(x),\sum_{n=0}^\infty c_n R^n = \lim_{x\to (a+R)^-} f(x),

so the sum of the boundary series can be computed as a one-sided limit of the closed form, provided convergence at the endpoint has been independently verified first.

Worked examples

Example 1: A series with different behavior at its two endpoints. Consider

f(x)=∑n=1∞(x−2)nn 3n.f(x) = \sum_{n=1}^\infty \frac{(x-2)^n}{n\, 3^n}.

Step 1 — radius of convergence. With cn=1n 3nc_n = \dfrac{1}{n\,3^n}, apply the ratio test to ∣cn+1(x−2)n+1∣/∣cn(x−2)n∣|c_{n+1}(x-2)^{n+1}| / |c_n (x-2)^n|:

∣cn+1cn∣∣x−2∣=nn+1⋅13 ∣x−2∣  ⟶  ∣x−2∣3(n→∞).\left| \frac{c_{n+1}}{c_n} \right| |x-2| = \frac{n}{n+1}\cdot\frac{1}{3}\,|x-2| \;\longrightarrow\; \frac{|x-2|}{3} \quad (n\to\infty).

The series converges absolutely when ∣x−2∣3<1\frac{|x-2|}{3} < 1, i.e. ∣x−2∣<3|x-2|<3, and diverges when ∣x−2∣>3|x-2|>3. So R=3R = 3, center a=2a=2, and the open interval of convergence contains (−1,5)(-1, 5). The two endpoints to check are x=−1x=-1 and x=5x=5. Step 2 — right endpoint x=5x=5. Here x−2=3x - 2 = 3, so the series becomes

∑n=1∞3nn 3n=∑n=1∞1n,\sum_{n=1}^\infty \frac{3^n}{n\,3^n} = \sum_{n=1}^\infty \frac{1}{n},

the harmonic series, which diverges (standard fact, e.g. via the integral test: ∫1∞dxx=∞\int_1^\infty \frac{dx}{x} = \infty, and 1n\frac1n compares directly to ∫nn+1dxx\int_n^{n+1}\frac{dx}{x}). So x=5x=5 is not in the interval of convergence. Step 3 — left endpoint x=−1x=-1. Here x−2=−3x-2=-3, so the series becomes

∑n=1∞(−3)nn 3n=∑n=1∞(−1)nn=−1+12−13+14−⋯ ,\sum_{n=1}^\infty \frac{(-3)^n}{n\,3^n} = \sum_{n=1}^\infty \frac{(-1)^n}{n} = -1 + \frac12 - \frac13 + \frac14 - \cdots,

the alternating harmonic series. Check the Alternating Series Test hypotheses for bn=1/nb_n = 1/n: (i) bn>0b_n>0 for all n≥1n\ge1; (ii) bnb_n is decreasing since 1n+1<1n\frac{1}{n+1}<\frac1n; (iii) bn→0b_n \to 0. All three hold, so the series converges (conditionally — it does not converge absolutely, since ∑1/n\sum 1/n diverges, as just shown). Conclusion. The interval of convergence is

I=[−1, 5),I = [-1,\ 5),

closed at the left endpoint, open at the right. This is a genuinely asymmetric outcome: dropping the same value of ∣x−2∣=3|x-2|=3 into the series with a ++ sign gives a divergent series (harmonic), while with a −- sign the alternation rescues convergence. By Abel's theorem applied at x=−1x=-1 (which lies in II), ff is continuous on (−1,5)(-1, 5) union the point x=−1x=-1 itself, i.e. on [−1,5)[-1,5), and

lim⁡x→−1+f(x)=∑n=1∞(−1)nn=−ln⁡2,\lim_{x \to -1^+} f(x) = \sum_{n=1}^\infty \frac{(-1)^n}{n} = -\ln 2,

using the known Taylor expansion ln⁡(1+t)=∑n=1∞(−1)n+1ntn\ln(1+t) = \sum_{n=1}^\infty \frac{(-1)^{n+1}}{n}t^n at t=1t=1 (itself justified by an application of this same endpoint analysis to the ln⁡(1+t)\ln(1+t) series, plus Abel's theorem).

Example 2: A series where a subtler comparison, not the alternating series test, is required. Consider

g(x)=∑n=1∞n24n (x+1)n.g(x) = \sum_{n=1}^\infty \frac{n^2}{4^n}\,(x+1)^n.

Step 1 — radius of convergence. With cn=n2/4nc_n = n^2/4^n, use the root test:

∣cn∣1/n∣x+1∣=(n1/n)24 ∣x+1∣  ⟶  14∣x+1∣(n→∞),|c_n|^{1/n} |x+1| = \frac{(n^{1/n})^2}{4}\,|x+1| \;\longrightarrow\; \frac{1}{4}|x+1| \quad (n\to\infty),

using the standard limit n1/n→1n^{1/n}\to 1. Convergence requires ∣x+1∣4<1\frac{|x+1|}{4}<1, i.e. ∣x+1∣<4|x+1|<4, so R=4R=4 centered at a=−1a=-1; open interval (−5,3)(-5,3). Endpoints to check: x=−5x=-5 and x=3x=3. Step 2 — right endpoint x=3x=3. Then x+1=4x+1=4, giving

∑n=1∞n24n⋅4n=∑n=1∞n2.\sum_{n=1}^\infty \frac{n^2}{4^n}\cdot 4^n = \sum_{n=1}^\infty n^2.

The general term n2→∞n^2 \to \infty, so the terms do not tend to 00; by the nn-th term test (divergence test — if ∑an\sum a_n converges then an→0a_n \to 0, so if an↛0a_n \not\to 0 the series diverges), the series diverges. Step 3 — left endpoint x=−5x=-5. Then x+1=−4x+1 = -4, giving

∑n=1∞n24n(−4)n=∑n=1∞n2(−1)n.\sum_{n=1}^\infty \frac{n^2}{4^n}(-4)^n = \sum_{n=1}^\infty n^2(-1)^n.

Again the general term ∣(−1)nn2∣=n2→∞≠0|{(-1)^n n^2}| = n^2 \to \infty \ne 0, so by the same nn-th term test the series diverges. (Note the alternating series test does not apply here at all: its hypothesis that bn=n2→0b_n = n^2 \to 0 fails outright, so there is no contradiction with convergence — the test's hypotheses simply are not met, and indeed the series diverges, consistent with, but not proved by, the alternating test.) Conclusion. The interval of convergence is

I=(−5, 3),I = (-5,\ 3),

open at both endpoints — the fourth possible outcome, distinct from Example 1. This example is instructive because it shows the endpoint analysis is not always "run the alternating series test and see." The correct first move at any endpoint is always to check whether the general term tends to 00; only if it does is the alternating series test (or a finer comparison/integral test) worth attempting.

Counterexample: why "converges at endpoint" cannot be weakened to "bounded partial sums only inside the interval"

CounterexampleDropping actual convergence at the endpoint breaks Abel's theorem

Abel's theorem requires that the series ∑cnRn\sum c_n R^n converges at the endpoint (hypothesis: existence of S=lim⁡NsNS = \lim_N s_N). It is tempting to think a weaker condition — say, merely that the partial sums sNs_N of ∑cnRn\sum c_n R^n stay bounded (without converging) — should still force some form of one-sided continuity or a well-defined boundary limit for ff. This fails. Take the geometric-type series (after normalizing a=0,R=1a=0, R=1)

f(x)=∑n=0∞(−1)nxn=11+x,∣x∣<1.f(x) = \sum_{n=0}^\infty (-1)^n x^n = \frac{1}{1+x}, \qquad |x|<1.

Radius of convergence is R=1R=1 (ratio test: ∣cn+1/cn∣=1|c_{n+1}/c_n|=1). At the right endpoint x=1x=1:

∑n=0∞(−1)n=1−1+1−1+⋯ ,\sum_{n=0}^\infty (-1)^n = 1 - 1 + 1 - 1 + \cdots,

whose partial sums are sN=1s_N = 1 if NN even, sN=0s_N=0 if NN odd — bounded (by 11) but not convergent, since the sequence sNs_N oscillates between two limit points 00 and 11 and has no single limit. The hypothesis of Abel's theorem (convergence of ∑cnRn\sum c_n R^n, not mere boundedness of partial sums) therefore fails at x=1x=1. And indeed the conclusion fails too: lim⁡x→1−f(x)=lim⁡x→1−11+x=12\lim_{x\to 1^-} f(x) = \lim_{x\to1^-}\frac{1}{1+x} = \frac12 does exist as a limit of the function, but there is no corresponding "sum of the series at x=1x=1" to which it could be equal, because that series simply has no sum — 12\frac12 is not the limit of any convergent numerical series ∑(−1)n\sum(-1)^n. The would-be conclusion "ff extends continuously to x=1x=1 with f(1)=∑(−1)nf(1)=\sum(-1)^n" is meaningless, since the right-hand side does not exist. This shows precisely which hypothesis is load-bearing: Abel's theorem is a statement about upgrading pointwise convergence at an endpoint to continuity up to that endpoint; it says nothing when the endpoint series diverges (even boundedly/oscillating), and no weaker replacement hypothesis of this kind suffices.

Scope, edge cases & common traps

SituationWhat to do
R=0R=0Series converges only at x=ax=a; no endpoint question.
R=∞R=\inftySeries converges for all real xx; no endpoint question.
Ratio/root test gives limit exactly 11 at some interior-looking xxThat xx is actually an endpoint of the interval; recompute RR correctly, don't force the test to decide it.
Endpoint series has terms not →0\to 0Diverges immediately by the nn-th term test; do not attempt the alternating series test.
Endpoint series alternates but terms are not eventually monotonicAlternating series test does not directly apply; use absolute convergence, Dirichlet's test, or grouping instead.
Endpoint series converges conditionallyIt lies in II; radius of absolute convergence is still RR, but II properly contains the open interval on that side.
Wanting to conclude endpoint sum equals a limit of a closed form before proving the series converges thereInvalid — Abel's theorem requires convergence as a hypothesis, established independently first.

The mathematician behind it

Origins · 1 minNiels Henrik Abel and how this was discovered
NH
Niels Henrik Abel
1802–1829 · Norway, early 19th-century analysis

Niels Henrik Abel worked in an era when the foundations of convergence and continuity were only just being placed on rigorous footing by Cauchy and others. Abel's 1826 paper on the binomial series was among the first to treat power series with full attention to the subtleties of convergence at the boundary of the interval — a level of care that was, at the time, unusual, since many contemporaries manipulated infinite series formally without checking convergence at all. Abel's short life (he died of tuberculosis at 26) produced foundational results across algebra (the unsolvability of the general quintic by radicals) and analysis (elliptic functions, the theorem on power series continuity that bears his name), work whose rigor anticipated the fully modern treatment of series that would follow later in the century.

History
- 1802 — Niels Henrik Abel born in Frøland, Norway.
- 1824 — Abel proves the impossibility of solving the general quintic equation by radicals.
- 1826 — Abel publishes his investigation of the binomial series ∑(mn)xn\sum \binom{m}{n}x^n, carefully treating convergence at x=±1x=\pm1; the boundary continuity result now called Abel's theorem appears in this circle of work.
- 1829 — Abel dies in Froland, Norway, shortly before a professorship offer from Berlin reaches him.
- Later 19th century — Abel's theorem becomes a standard tool in the rigorous theory of power series developed by Weierstrass and successors, and a template for "boundary behavior" questions throughout analysis.
Why can't I just plug x=a+Rx = a+R into the ratio test formula and see what happens?

Because at x=a+Rx=a+R the ratio test's limit is, by construction, exactly 11 (that is how RR was defined in the first place: the boundary between the ratio being <1<1 and >1>1). The ratio test's own statement explicitly excludes the case where the limit equals 11, declaring the test inconclusive there. Plugging the endpoint value into the ratio test simply reproduces this inconclusive case every single time, for every power series, with no exceptions — so it supplies zero information. The only way forward is to substitute the endpoint value into the original series itself, producing a genuine numerical series (no longer depending on xx), and then apply a convergence test suited to that specific numerical series: comparison, alternating series, integral test, or direct manipulation.

Is it possible for a power series to diverge at both endpoints even though it converges absolutely just barely inside?

Yes, and Example 2 in this article is exactly such a case: ∑n2(x+1)n/4n\sum n^2(x+1)^n/4^n has radius 44 and diverges at both x=3x=3 and x=−5x=-5, because the coefficients n2/4nn^2/4^n decay only like a polynomial correction to the exponential rate, which is not enough to overcome the n2n^2 growth once ∣x+1∣=4|x+1|=4 exactly cancels the exponential factor. In general, whenever ∣cn∣Rn↛0|c_n| R^n \not\to 0, both endpoints diverge automatically by the nn-th term test, regardless of any alternation in sign.

Does Abel's theorem tell us whether the endpoint series converges?

No — that is a common misreading. Abel's theorem assumes the endpoint series converges (that is its hypothesis) and then concludes something about continuity of the function up to that point. It is not a convergence test; it never decides whether ∑cnRn\sum c_n R^n converges. You must first establish convergence at the endpoint using an ordinary series test (comparison, alternating series, integral test, etc.), and only afterward may you invoke Abel's theorem to relate the resulting sum to a limit of the function.

If the series converges absolutely at one endpoint, does it automatically converge at the other endpoint too?

No. Absolute convergence at x=a+Rx=a+R means ∑∣cn∣Rn<∞\sum |c_n| R^n < \infty, and since ∣cn(−R)n∣=∣cn∣Rn|c_n(-R)^n| = |c_n|R^n as well, this does force absolute (hence ordinary) convergence at x=a−Rx=a-R too — so if one endpoint converges absolutely, the other does as well, giving I=[a−R,a+R]I=[a-R,a+R]. However, if the endpoint converges only conditionally (not absolutely), this argument breaks down, since conditional convergence depends on the specific sign pattern created by (±R)n(\pm R)^n, and the other endpoint can behave completely differently, exactly as in Example 1.

Can I use L'Hôpital's rule or Taylor's theorem shortcuts to skip endpoint analysis?

No shortcut of this kind exists for deciding convergence of the endpoint series itself; those tools apply to limits and derivatives of functions, not to determining whether an infinite sum of specific numbers converges. Endpoint convergence is fundamentally a question about a numerical series ∑an\sum a_n, and must be settled using series convergence tests: comparison, limit comparison, the alternating series test, the integral test, Dirichlet's test, or direct estimation of ∣an∣|a_n|'s asymptotic size. Abel's theorem then connects a known-convergent endpoint sum to the function's boundary limit, but it does not replace the need to first prove that convergence.

What if the two endpoint series both involve factorials or n-th roots making standard tests awkward?

Use asymptotic comparison: estimate cnRnc_n R^n up to a bounded multiplicative factor using Stirling's approximation for factorials, or standard limits like n1/n→1n^{1/n}\to1, and then compare to a pp-series ∑1/np\sum 1/n^p or geometric series whose convergence is already known. This is exactly the technique used implicitly in both worked examples: reducing an unfamiliar series to a recognizable comparison series (the harmonic series, the alternating harmonic series, or ∑n2\sum n^2) via direct algebraic simplification of cnRnc_n R^n.

Before this